A machine learning-based prediction of crystal orientations for multicrystalline materials

Abstract

We established a rapid, low-cost, and accurate technique to measure crystallographic orientations in multicrystalline materials by optical images and machine learning. A long short-term memory neural network was trained with pairs of light reflection patterns and the correct orientations of each grain, successfully predicting orientation with an error median of 8.61°. The model was improved by diverse data taken from various incident light angles and by data augmentation. When trained on different incident angles, the model was capable of estimating different orientations. This is related to the geometrical configuration of the incident light angles and surface facets of the crystal. The failure in certain orientations is thought to be complemented by supplementary data taken from different incident angles. Combining data from multiple incident angles, we acquired an error median of 4.35°. Data augmentation was successfully performed, reducing error by an additional 35%. This technique can provide the crystallographic orientations of a 15 × 15 cm2 sized wafer in less than 8 min, while baseline techniques such as electron backscatter diffraction and Laue scanner may take more than 10 h. The rapid and accurate measurement can accelerate data collection for full-sized ingots, helping us gain a comprehensive understanding of crystal growth. We believe that our technique will contribute to controlling crystalline structure for the fabrication of high-performance materials.